What the October price data actually shows
TechSpot published its fourth-quarter GPU pricing survey on November 2, 2026. It is not an average price per model. It tracks the single lowest in-stock price for each card in ten countries, confirms that the unit is actually purchasable at that price, and compares it against the launch MSRP converted into local currency using exchange rates and tax. October came in 11 percent above September and 35 percent above August. Across all models and all ten countries the lowest shelf price averaged 44 percent above MSRP.
The average is not the interesting part. The interesting part is the spread between three cards that carry the same memory.
| Model | Memory | Launch MSRP | Lowest US price, Oct 2026 | Above MSRP |
|---|---|---|---|---|
| RTX 5060 Ti 16GB | 16 GB GDDR7 | $429 | $790 | $361 |
| RTX 5070 Ti | 16 GB GDDR7 | $749 | $1,250 | $501 |
| RTX 5080 | 16 GB GDDR7 | $999 | $2,200 | $1,201 |
Memory cost is a function of capacity and grade. Three cards with 16 GB of the same standard carry a near-identical memory line in the bill of materials, so the $140 difference between the first two markups and the $840 difference between the first and the third cannot be memory. The same pattern holds on the ten-country averages: roughly $300 above MSRP for the 5060 Ti 16GB, roughly $450 for the 5070 Ti, roughly $770 for the 5080.
That does not mean every dollar of the gap is pure markup. The die, the board, the power circuitry and the cooling differ between these three cards, and a larger die costs more to fabricate and cool. What the comparison establishes is that the memory line is fixed across all three, so whatever drives the difference sits on the other side of the bill of materials. TechSpot states this directly for the 5080: the cost of 16 GB of GDDR7 has not risen by $700, and it names price gouging, AI-driven demand and the scarcity of the RTX 5090 as candidate causes. The size of each is unknown. One caveat on the 5080 figure: at the time of the survey Newegg in the US had no direct stock of that card, only third-party sellers at inflated prices, so the $2,200 may reflect marketplace pricing rather than a price the channel is honoring.
What memory actually costs on a card
The memory market is genuinely tight. TrendForce’s September 30 research says DRAM suppliers continue to prioritize advanced-process capacity for high-performance server products through 4Q26, with conventional DRAM contract prices projected up 10 to 15 percent quarter over quarter. In graphics DRAM specifically, suppliers are scaling back GDDR6 production and gradually exiting the market, while GDDR7 output stays limited because advanced-process capacity is being allocated to server DRAM. TrendForce also notes that GDDR7 demand is driven primarily by AI chips rather than by PC builds, so a gaming card is no longer the deciding factor in where that capacity goes.
Per-unit prices are the missing piece. TrendForce publishes no per-module price for either standard. The published figure comes from DRAMeXchange via a September 9 industry report: the average transaction price of an 8 Gbit (1 GB) GDDR6 module went from $2.80 in late October 2025 to $8.40. Eight of those chips, which is what an 8 GB card uses, went from $22.40 to $67.20 in bill-of-materials cost. That is a 3x change on a single line item, and it is the only per-unit number anyone has published.
AMD reportedly raised the GDDR6 price it supplies to its board partners effective October 1, 2026. The increase was reported by a Chinese industry site and picked up by BenchLife. AMD has not confirmed it, and the size was not disclosed. What matters is the assumption it breaks. The long-standing idea that choosing the older memory standard keeps a card cheap no longer holds, because the older standard is being produced less.
The 8 GB to 16 GB gap is the cleaner test
Comparing two variants of the same GPU removes the die, the board and the cooler from the comparison. Only capacity differs.
| Model | Memory | Launch MSRP | Lowest US price, Oct 2026 | Above MSRP |
|---|---|---|---|---|
| RTX 5060 Ti 8GB | 8 GB GDDR7 | $379 | $530 | $151 |
| RTX 5060 Ti 16GB | 16 GB GDDR7 | $429 | $790 | $361 |
| RX 9060 XT 8GB | 8 GB GDDR6 | $299 | $500 | $201 |
| RX 9060 XT 16GB | 16 GB GDDR6 | $349 | $620 | $271 |
Both pairs launched with a $50 difference between the 8 GB and 16 GB variants. In October the lowest US prices put that gap at $260 for the GeForce pair and $120 for the Radeon pair.
Using the only published per-unit figure, eight additional 8 Gbit GDDR6 chips cost $67.20 at current prices. Taking that as a proxy, the arithmetic puts roughly $193 of the GeForce gap and roughly $53 of the Radeon gap outside the memory line. The proxy is weak for the GeForce pair, because those cards use GDDR7 and no per-unit GDDR7 price is published. The Radeon pair is the case where the published number actually applies, and there the memory cost accounts for most of the widened gap. On the GeForce side it accounts for about a quarter of it.
Here is the part that matters for anyone buying a card to run a model. Both variants of the RTX 5060 Ti run a 128-bit interface at 28 Gbps, which is 448 GB/s. Both variants of the RX 9060 XT run a 128-bit interface at 20 Gbps, which is up to 320 GB/s. The 16 GB variant does not get a wider bus. It gets more chips on the same bus. Paying the extra $260 or $120 buys capacity, not bandwidth, and capacity is not a speed feature. It is a fit feature. It decides whether the weights and the key-value cache fit on the card at all. Once they fit, the bus width is what sets the rate, and the bus width is identical.
Bandwidth per dollar is the number that governs a decode loop
Decode is a memory-read problem. A single token requires one pass over the model weights. For a 7B model quantized to 4 bits, the weights are about 4.2 GB, so the upper bound on decode rate is the memory bandwidth divided by 4.2. That is arithmetic, not a measurement, and it is an upper bound: it ignores the key-value read, the attention and projection overhead, and the fact that no real card sustains its theoretical peak.
| Model | Bus | Bandwidth | Lowest US price, Oct 2026 | GB/s per dollar | Arithmetic ceiling, 7B Q4 |
|---|---|---|---|---|---|
| Arc B570 | 160-bit | 380 GB/s | $290 | 1.31 | 90 tok/s |
| Arc B580 | 192-bit | 456 GB/s | $350 | 1.30 | 109 tok/s |
| RTX 5060 Ti 8GB | 128-bit | 448 GB/s | $530 | 0.85 | 107 tok/s |
| RTX 5060 Ti 16GB | 128-bit | 448 GB/s | $790 | 0.57 | 107 tok/s |
| RX 9060 XT 16GB | 128-bit | 320 GB/s | $620 | 0.52 | 76 tok/s |
| RX 9070 | 256-bit | 640 GB/s | $700 | 0.91 | 152 tok/s |
| RTX 5070 | 192-bit | 672 GB/s | $900 | 0.75 | 160 tok/s |
| RTX 5070 Ti | 256-bit | 896 GB/s | $1,250 | 0.72 | 213 tok/s |
| RTX 5080 | 256-bit | 960 GB/s | $2,200 | 0.44 | 229 tok/s |
| RTX 5090 | 512-bit | 1,792 GB/s | $9,400 | 0.19 | 427 tok/s |
At launch MSRP the ordering was different. A $249 Arc B580 delivered 1.83 GB/s per dollar, a $249 Arc B570 delivered 1.74, and the RTX 5090 delivered 0.90. The pricing structure has moved the value toward the narrow-bus cards, because those are the ones whose prices did not move as far. The 5090 is now 370 percent above its MSRP in the US, up from 175 percent in February. The 5060 Ti 8GB is 40 percent above, the lowest inflation of any GeForce card in the survey.
Two cards break the pattern in opposite directions. The RX 9070 is a 256-bit card at 640 GB/s for $700, which is 0.91 GB/s per dollar, better than the 5060 Ti 16GB at 0.57 and better than the 5080 at 0.44. It carries the same 16 GB as the 5060 Ti 16GB on a bus twice as wide. The Arc B580 is 12 GB on a 192-bit bus at 456 GB/s for $350. Intel narrowed the bus from the 256-bit A580 to compensate with a faster 19 Gbps memory clock, which still came out 11 percent below the previous generation. For a model that fits in 12 GB, that card is the cheapest path to a decode rate above 100 tokens per second in this table.
What this means for a local build
Three things follow from the data, and none of them are about gaming.
First, capacity and bandwidth are separate purchases. If a model fits on an 8 GB card, the 16 GB variant of the same GPU runs it at the same rate. The extra money buys context headroom, not a faster token. If the model does not fit, the card stops working on the GPU and starts paging, and that path is much worse than the headline numbers: the RTX 5060 Ti is limited to eight PCIe lanes, which is 64 GB/s on PCIe 5.0, 32 GB/s on 4.0 and 16 GB/s on 3.0, against a system memory peak of roughly 96 GB/s on dual-channel DDR5-6000. The offload tier is where the throughput goes, and its block-size behavior is covered in the DRAM-less SSD queue depth test and the block-size comparison.
Second, read the bus width, not the memory type. A 128-bit card at 28 Gbps and a 128-bit card at 20 Gbps are 448 GB/s and 320 GB/s. The GDDR7 label carries a price premium: the extra-capacity markup is $210 on the GeForce pair against $70 on the Radeon pair, and it buys 128 GB/s on the same bus width. Whether that is worth it depends entirely on whether the model fits. The same reasoning on a smaller board produced the ceiling on local decode from a 273 GB/s edge module.
Third, the price signal is not a demand signal. Nvidia’s Graphics segment reported $7.922 billion for the quarter ended July 26, up 46 percent year over year, with operating income of $3.899 billion up 74 percent. Nvidia does not disclose GeForce revenue, unit volume or average selling price separately, so there is no way to tell whether that came from price or volume. AMD’s Gaming segment fell 31 percent to $779 million, but the stated cause was lower semi-custom console sales, so it says nothing about Radeon supply. The only direct supply statement is Nvidia’s consumer computing senior vice president Jeff Fisher, who said after the RTX Spark Launch Party in Taipei that 5090 production continues while acknowledging capacity is “indeed somewhat constrained.” The card has effectively disappeared from US retail regardless of what the production line is doing.
The practical decision
For a build whose job is running models, the buying logic is: pick the widest bus you can get at a price you can pay, then check whether the model fits on it. That is the same framework as bandwidth per dollar, and the October data moved it. The 32 GB tier is where the pricing broke completely, which is what the GB202 allocation covers. Below that tier the cards are still comparable on a per-dollar basis, and the comparison is between bus widths, not between memory standards.
What is actually available at those prices right now, as of November 4, 2026:
- ASUS Dual RTX 5060 Ti 8GB at about $548. The narrow-bus card at the lowest inflation in the survey. Fine for a 7B model that fits, and it runs at the same rate as its 16 GB sibling.
- PNY RTX 5060 Ti 16GB at about $770. Same 448 GB/s, more headroom for context. The $222 difference buys fit, not speed.
- ASUS Dual RX 9060 XT 16GB at about $726. 320 GB/s on a 128-bit bus with 32 MB of Infinity Cache. Cheaper than the GeForce 16 GB variant and slower on a decode loop.
- ASUS Prime RX 9070 XT 16GB at about $1,051. The 256-bit option. This is the card in the list where the price is defensible on bandwidth grounds, because the bus is twice as wide as the 9060 XT’s.
Prices in this article are the lowest verified in-stock figures from TechSpot’s October survey for the model-level comparison, and live Amazon listings for the specific cards above. They are approximate and they move. The ratios, the bus widths and the arithmetic ceilings are the parts that do not move, and those are the parts worth acting on.

